W&B (Weights & Biases) Integration

What it does

This skill lets an agent interact programmatically with Weights & Biases (W&B) using the Membrane CLI. It provides guidance and commands for authenticating, creating or ensuring connections, discovering available actions (runs, projects, artifacts), and running actions to query or update W&B resources. The primary value is making W&B operations safe and repeatable from an agent context: auth is delegated to Membrane, CLI calls return JSON, and common workflows (search actions, run actions, proxy requests) are documented.

When to use it

Use this skill when a user asks the agent to inspect experiment metadata, list or compare runs, fetch artifacts, push metrics, or automate repeatable W&B workflows (for example: "show latest runs for project X", "download artifact Y", "compare metrics between runs"). It's also appropriate when the agent needs headless authentication instructions or to orchestrate actions via Membrane in CI or automated environments.

What's included

• Scripts: none bundled in-skill, but the instructions reference using the Membrane CLI (membrane) as the primary tool.
• References: links to official W&B docs and Membrane CLI usage are provided in the skill body.
• Instructions: step-by-step guidance for installing the CLI, authenticating (interactive and headless flows), ensuring a connection, polling for connection readiness, searching for actions, running actions with JSON input, and proxying raw HTTP requests through Membrane. Best-practice notes on preferring Membrane for auth/pagination and using --json for machine-readable output are included.

Compatible agents

Designed for agents that can run shell commands and call CLIs (Claude Code, OpenClaw-style agents, Codex/Copilot-style environments). The skill assumes the Membrane CLI is available and network access to W&B.